Anthropic
Anthropic's balanced workhorse — the everyday coding and agentic model.
Sonnet trades a little of Opus's ceiling for speed and cost, which is exactly the niche mid-size open-weight models fill best. These are the community-regarded closest local models by capability class:
These are community-regarded closest open-weight models by capability class — an editorial judgement, not a benchmark ranking. Sizes and hardware/speed figures are derived from the catalog and the same fit engine used across the site, judged at a 8,192-token context. Speeds are estimates, not measurements.
Mid-size open coding models that run comfortably on a single 24 GB desktop GPU are the closest practical stand-ins for Sonnet's day-to-day coding.
| Local model | Size | Hardware to run it | Est. speed |
|---|---|---|---|
| Qwen3-Coder-30B-A3B-Instruct Qwen tags: Coding, Long context | 30.5B (MoE) | Runs fully on 20 GB (RX 7900 XT) at IQ4_XS Runs fully on GPU @ IQ4_XS · 8K ctx | 184–246 tok/s (estimate) |
For Sonnet-style balanced generalist and tool-use work, these mid-size open instruct models are the community's usual local choices.
| Local model | Size | Hardware to run it | Est. speed |
|---|---|---|---|
| Qwen3-32B Qwen | 32.8B | Runs fully on 24 GB (RTX 4090) at Q4_K_M Runs fully on GPU @ Q4_K_M · 8K ctx | 28–37 tok/s (estimate) |
| Qwen3-30B-A3B-Instruct-2507 Qwen tags: Long context | 30.5B (MoE) | Runs fully on 24 GB (RTX 4090) at Q4_K_M Runs fully on GPU @ Q4_K_M · 8K ctx | 213–284 tok/s (estimate) |
| Mistral-Small-3.2-24B-Instruct-2506 mistralai | 24B | Runs fully on 16 GB (RTX 5060 Ti 16GB) at Q4_K_M Runs fully on GPU @ Q4_K_M · 8K ctx | 19–25 tok/s (estimate) |
| gpt-oss-20b openai | 21.5B (MoE) | Runs fully on 16 GB (RTX 5060 Ti 16GB) at F16 Runs fully on GPU @ F16 · 8K ctx | 77–103 tok/s (estimate) |
"Hardware to run it" is the smallest single GPU in our roster that runs the model's best-fitting quant fully on GPU at this context, computed by the same fit engine used sitewide. "Est. speed" is a modelled generation-throughput range labelled estimate (D8), never a measurement. Equivalence is a conservative editorial call and no benchmark scores are implied.